Multiple Imputation – How to Apply Inverse Probability Weighting for Multiple Treatments

multinomial logitmultiple-imputationpropensity-scores

I am analyzing observational study data. My predictor variable is tg4 with 4 categories (0,1,2,3) and my response variable is dm (0,1).
my analysis step as follow:

  1. I use mice package to impute missing data as follow
    chi<-chi %>% mice(seed = 123, print = FALSE, m=5)
  2. Covariate balance assessment using standardized mean differences (SMD) (I tried to use tableone package) but it does not work because I have 5 imputed datasets.
  3. calculating propensity score using multinomial logistic regression.
  4. calculating weight
  5. calculating pooled hazard ratio for all dataset
suvtime<-Surv(time=chi$time,event = chi$dm)
fit<- coxph(suvtime~tg4,data=chi, weights =weights)

Thank you for your help.

Best Answer

There is a set of packages you can use for this. The MatchThem package performs matching and IPW in multiply imputed datasets; you can supply it with a mids objects from mice and it will perform the weighting in all of the imputed datasets. It calls the weightit() function from the WeightIt package to do the weighting. You can then use the bal.tab() function in cobalt to assess balance on the imputed datasets. Once you have achieved balance, you can fit your outcome model using the with() function in MatchThem.

I recommend you 1) look at the WeightIt package and see its capabilities for multi-category treatments (there are many options beyond multinomial logistic regression); 2) look at the MatchThem package for applying WeightIt functionality to multiply imputed datasets and estimating treatment effects pooling across the datasets; and 3) look at the cobalt package for assessing balance generally and specifically with multiply imputed data and multi-category treatments. I am the author of all three of these packages, and they are designed to work seamlessly together specifically for this type of analysis.

Related Question